Bundle: run-scoped provider permanent-failure fall-through (#1936)#1961
Open
Serhan-Asad wants to merge 4 commits into
Open
Bundle: run-scoped provider permanent-failure fall-through (#1936)#1961Serhan-Asad wants to merge 4 commits into
Serhan-Asad wants to merge 4 commits into
Conversation
Agentic provider selection already fell through to the next provider on a
permanent error *within a single run_agentic_task call*, but each call
rebuilt its candidate list from scratch. A permanently-dead provider (e.g.
Codex whose staged ChatGPT auth is present but expired) was therefore
re-attempted on every step of a multi-step workflow, and a call whose only
feasible candidate was the dead provider had nowhere to fall through to —
so cloud one-session sync died on codex's 401 despite healthy anthropic/
google credentials in the same run.
Add a run-scoped permanent-failure registry in the single provider-selection
source of truth (run_agentic_task):
- mark_provider_permanently_failed(provider, classification) records a
provider when _classify_permanent_error fires; get_disabled_providers()
exposes the {provider: classification} map; reset_disabled_providers()
clears it. The registry is mirrored into the PDD_AGENTIC_DISABLED_PROVIDERS
env var so re-entrant `pdd` subprocesses inherit the same skip-list.
- run_agentic_task drops disabled providers from candidates AFTER the
PDD_AGENTIC_PROVIDER hard filter and BEFORE the single_provider_attempt
truncation, so a known-dead first provider no longer consumes the single
attempt and later steps skip it instead of re-paying its timeout.
- When every feasible provider was already disabled this run, fail fast with
an aggregated "All agent providers failed permanently earlier this run:
<provider>: <class>; ..." message instead of re-burning or the generic
"no providers available" line.
The registry is subordinate to PDD_AGENTIC_PROVIDER: it only ever removes a
provider that proved dead, never adds one, so a staged-auth hint cannot
re-pin a provider once its auth is known bad. No behavior change when the
first provider works.
Updates the owning prompt (agentic_common_python.prompt) with the new
exports and the registry requirement, and adds regression tests plus
conftest isolation for the new env var / registry.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rough Run-scoped provider permanent-failure fall-through (#1936). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Collaborator
Author
Staging E2E evidence (2026-07-09, staging1, stack/e2e-staging-v2 = 288d053)Validated on every lane, all day, against the genuinely-dead codex fleet:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Bundle: run-scoped provider permanent-failure fall-through
Off current
main, adds run-scoped fall-through so a provider family that has reported a permanent failure once during a run is skipped for the rest of that run instead of being retried into the same wall.Design
A per-run registry in
agentic_commonrecords provider families that have hit a permanent (non-transient) failure — auth/credential-limit and other terminal classifications — and short-circuits further dispatch to that family for the remainder of the run, falling through to the next eligible provider. The registry is run-scoped (not process-global), so it never leaks across independent runs, and transient failures are deliberately excluded so normal retry/backoff is unaffected.Root-cause correction
The initial framing treated this as a routing/registry gap; the corrected root cause — that the permanent-failure signal was observed but not made authoritative for subsequent dispatch within the run — is documented at:
#1936 (comment)
Closes #1936
Validation
tests/test_agentic_common.py— 525 passed on the bundle branch.🤖 Generated with Claude Code